SimCLR self-supervised representation learning
SimCLR with NT-Xent loss on ResNet-18/50 over CIFAR-10: ~96.5% top-5 test accuracy using only 10% of labels.
- Status
- Research
- Period
- 2025
- Area
- Deep learning
- Code
- Code public
- Stack
- Python
- PyTorch Lightning
- ResNet-18/50
- CIFAR-10
- Top-5 test accuracy
- ~96.5%
- Labels used
- 10%
- Image
- Two augmented views
- ResNet encoder
- Projection head
- NT-Xent loss
Problem
Labels are expensive. Learn useful image representations first, then use only a small fraction of labels.
Approach
I implemented SimCLR contrastive learning with the NT-Xent loss on ResNet-18 and ResNet-50 backbones over CIFAR-10.
Results
~96.5% top-5 test accuracy using only 10% of labels.